1048046 - 5.7% regression in octane-crypto on MacOSX 32bit

Status: RESOLVED FIXED

(Core :: JavaScript Engine: JIT, defect)

Description

[Hannes Verschore [:h4writer]]

AWFY is reporting a 5.7% regression in octane-crypto on MacOSX 32bit
http://arewefastyet.com/#machine=11&view=single&suite=octane&subtest=Crypto&start=1406547180&end=1406555257

Possibly bug 1040194?

Comment 1

[Nicolas B. Pierron [:nbp]]

We can revert the patch if needed, as this is protecting us against some escape analysis cases, which is disabled by default.

I will submit a fix today.  The problem here is that we are no longer eliminating resume point operands which are completely dead in baseline, if we can recover the instruction.  Recovering an instruction is costly compared to deleting the instruction and not having to recover it at all.  This spilling overhead is what is seen here as the 5.7% regression.

Comment 2

[Nicolas B. Pierron [:nbp]]

Attachment: Remove non-observed instructions, even if they might be recovered on bailout. r=

Comment 3

[Nicolas B. Pierron [:nbp]]

https://hg.mozilla.org/integration/mozilla-inbound/rev/c835743934c3

Comment 4

[Ryan VanderMeulen [:RyanVM]]

https://hg.mozilla.org/mozilla-central/rev/c835743934c3

Comment 5

[Hannes Verschore [:h4writer]]

http://arewefastyet.com/#machine=11&view=single&suite=octane&subtest=Crypto&start=1407138173&end=1407199231